Bronchopulmonary dysplasia (BPD) is a multifactorial neonatal lung disease that remains difficult to predict and biologically characterize. Recent years have seen multiple omics studies aimed at discovering molecular biomarkers and pathways that reflect disease heterogeneity and pathogenesis. The study summarized here applied a sparse machine learning approach to integrate multiple cord blood omics platforms with the goal of identifying reliable, parsimonious biomarker signatures and the biological processes they represent.
The investigators set out to evaluate the performance of three distinct omics platforms—metabolomics, proteomics, and adductomics—when integrated through a sparse machine learning pipeline. The pipeline incorporated Least Absolute Shrinkage and Selection Operator (LASSO) regression together with the sparse, reliable and adaptive biomarker identification method (Stabl) to assess predictive power for BPD and to select stable, low-dimensional biomarker sets.
Analyses used a well-characterized birth cohort totaling 217 infants. The cohort composition reported in the source included 52 term infants and 165 extremely preterm infants (<28 weeks gestation). Among the preterm subgroup, 82 infants developed BPD and 35 experienced severe BPD or death. The preprint reports that these clinical outcome categories were the focus for model development and validation within the cohort.
Cord blood specimens were profiled across three omics platforms: metabolomics, proteomics, and adductomics. In total, sparse multivariable modeling evaluated approximately 45,000 molecular features measured across the 217 infants. These high-dimensional data were the input for the sparse selection and predictive modeling pipeline described below.
The analytic strategy integrated two complementary sparse modeling approaches. LASSO regression was applied to perform shrinkage and selection among the many candidate features. In parallel or in combination, the authors used Stabl, a method designed for sparse, reliable, and adaptive biomarker identification, to prioritize features that are both predictive and stable across resampling. The pipeline was used to evaluate diagnostic discrimination for preterm birth and for BPD outcomes, and to produce parsimonious biomarker signatures.
When classifying preterm versus term birth using the high-dimensional feature set, both LASSO and Stabl identified a signature that achieved perfect discrimination (AUROC = 1.0; p < 0.001) within the study dataset. Restricting analyses to the preterm group, sparse multivariable modeling demonstrated strong predictive performance for severe BPD, with reported AUROC = 0.83 (p = 0.005). These results indicate robust discrimination in the cohort for the outcomes evaluated, as reported in the preprint.
Using Stabl, the authors identified a subset of 12 biomarkers that collectively predicted grade III BPD with good performance (AUROC = 0.76; p = 0.03). The selected panel comprised 2 adducts, 3 proteins, and 7 metabolites. The source reports these molecular classes but does not list individual analytes or provide their identities in the abstract; details and feature names may be available in the full preprint or supplementary materials referenced by the authors.
Across the three omics platforms, the selected biomarkers and pathway analyses pointed to dysregulated biological processes relevant to BPD pathogenesis. Specifically, the findings implicated perturbations in innate and adaptive immune responses, altered metabolic programming, and evidence of oxidative stress. These pathway-level signals suggest multi-system contributions to BPD and align with the multifactorial nature of the disease as characterized in the study.
The authors conclude that a sparse machine learning pipeline combining LASSO and Stabl can serve as a complementary approach for identifying novel pathways and low-dimensional biomarker panels for multifactorial conditions such as BPD. Within this cohort, the approach yielded a perfect signature for preterm birth, strong discrimination for severe BPD among preterm infants, and a 12-feature signature predictive of grade III BPD. The work illustrates how integrating multi-omic cord blood data can reveal candidate biomarkers and mechanistic pathways that warrant further validation.
The preprint lists NIH funding support with grant identifiers R01HL139798 and R21HD100831. The authors declared no competing interests. The abstract references supplementary material and data/code links; however, the abstract itself does not provide full feature identities, external validation results, or detailed limitations. Interested readers should consult the full preprint and supplementary files for complete methods, feature lists, and validation details.